FACET-0 autonomous rollout

PINE Lab · Nanyang Technological University · Singapore

Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation

ManuFacet-1K, a force-synchronized precision-assembly dataset — and FACET-0, the foundation model it trains: reasoning, acting, and correcting at 0.5 mm.

▶ Watch the system work 🤗 FACET-0 model ↗ 🤗 ManuFacet-1K dataset ↗ Paper arXiv Code

Haoyuan Deng^*, Haichao Liu^*, Wenkai Guo^*, Yuan Ling, Zaijia Yang, Yuanjiang Xue, Haosheng Sun, Liangzi Wang, and Ziwei Wang^† PINE Lab, Nanyang Technological University, Singapore * Equal contribution · † Corresponding author

0 hdemonstrations + rollouts · ManuFacet-1K

0 mmaction precision · fine assembly regime

0high-precision assembly tasks

0embodiments · one shared schema

0collection sites · live network

The problem

Our world is built on assemblies

Assembly, insertion, fastening, alignment — the operations that decide quality and yield are still labor-intensive and brittle to automate.

Aero-engine assembly≈18,000parts in a single Trent engine

Semiconductor precision electronics assembly

Semiconductor & electronics±5–25 µmchip placement / bonding precision

Assembly · inspection · loading2.1Munfilled US manufacturing jobs by 2030

Sources: Rolls-Royce · Palomar Technologies / Semiconductor Digest · Deloitte × The Manufacturing Institute, 2021.

Status quo

Pre-programmed automation

Fixed trajectories, fixed fixtures

Highly efficient at a single task
Can't adapt to part variation or layout changes
Painful setup — time, money, expert teaching

Status quo

Human labor

Skilled operators on the line

Flexible and adaptive
Structural labor shortage
Painful runtime — fatigue, mental health, safety

Our approach

Data-driven manufacturing intelligence

Foundation models trained on real industrial data

Adapts across parts, layouts, and product batches
Industrial-grade precision and repeatability
Continually improves while deployed on the line

The bottleneck

Scaling ≠ precision — the gap is manufacturing data

Under 1% of public robot data touches manufacturing assembly — and scale buys generalization, not the contact judgment, force correction, and recovery that ±0.1 mm insertion demands. Crossing that gap takes a dataset built for contact.

General-purpose robot data · Open X-Embodiment, AgiBot World…~11,000 h

Manufacturing assembly in public data · REASSEMBLE + FMB44 h

ManuFacet-1K · force-synchronized precision assembly1,000 h

Public dataset hours: Open X-Embodiment ~7–11k h · AgiBot World 2.9k h · RoboCasa365 2k h · REASSEMBLE 13 h · FMB 31 h.

Demonstrations

Four assemblies, fully autonomous

One uncut evaluation run on a real Dell workstation — RAM, CPU, GPU, and disk, back to back.

One take — all four tasks, no cuts · FACET-08×

Four task results — end to end

Third-person view · full FACET-0 episodes

RAM insertion ★★★TIMELAPSE

CPU placement ★★★★★TIMELAPSE

GPU installation ★★★★TIMELAPSE

Disk mounting ★★★★TIMELAPSE

Key skills — up close

The contact-rich moments that decide each task

RAM · DIMM seating & latch

CPU · socket-level placement

GPU · PCIe alignment & press

Disk · caddy engagement

The dataset

ManuFacet-1K — one dataset, three embodiments

1,000 hours of force-synchronized precision assembly — one schema, one aligned action space, collected live across two sites.

LIVE

641.911h trainable

2 collection sites · one schema, one QA gate

25,153curated episodes

+28.3 hlast 24 h

90%QA pass rate

CPU239.724 h

RAM140.548 h

Disk150.280 h

GPU111.360 h

Force-synchronized · wrist F/T on every frame One clock · 160 Hz state, 15 Hz vision Language · subtask semantic tags Recovery-rich · failure + recovery segments kept 0.5 mm · action-precision regime